How DreamTavern works

On DreamTavern, you can talk to the characters as if they are real people. Understanding basic details on how the characters generate their responses will give a more nuanced understanding of how to create and edit your experience.

What is a Large Language Model (LLM)?

An LLM is a type of artificial intelligence that understands and generates human-like text. It's like having a robot that has read a vast amount of books, articles, and websites and can use that knowledge to write or complete sentences just like a human would.

The Basics of Language Models

Imagine if you had read everything ever written in your language. You'd be pretty good at guessing what comes next in a sentence, right? That's essentially what an LLM does. It uses statistics and probabilities to predict the next word in a sentence based on the words that come before it.
DreamTavern provides you with access to the state-of-the-art language models that are either developed by independent developers around the world or private companies like OpenAI. Each model has a different flavor, with different strengths and weaknesses.

How does the AI generate text?

Step-by-step text generation

Once trained, an LLM generates text one word (or piece of a word) at a time:
  1. 1.
    Starting Point: It begins with a prompt, which could be as simple as "The weather today is..."
  2. 2.
    Prediction: Based on the prompt, the LLM predicts several possible next words.
  3. 3.
    Selection: The LLM then selects the most likely next word based on what it has learned.
  4. 4.
    Repetition: This process repeats for each new word until a full sentence or paragraph is generated.

Coherence and relevance

The challenge for an LLM is not just to generate grammatically correct text, but also to make sure it is coherent (makes sense) and relevant to the topic. To do this, LLMs use context from the provided text, which is similar to keeping track of the topic during a conversation.
To summarize:
  • Text generation starts from a text prompt.
  • LLMs attempt to "auto-complete" this text in a way that's coherent.
  • The model that you use has an opinionated way of how it "likes" to be prompted.
  • Depending on the prompt, the text generation can vary widely.

How does this apply to DreamTavern?

Every text generation starts from a prompt template. DreamTavern then inserts information about the characters in this prompt alongside other conversation settings. This finalized prompt then gets fed to the LLM, which will generate a response.
Because prompting LLMs can have such a powerful impact on the end result, DreamTavern attempts to offer creators a wide array of tools and settings to customize their experience. The next section provides an overview of this process.
Last modified 2mo ago